Artificial Intelligence 171 (2007) 619 641 www.elsevier.com/locate/artint Argumentation in artificial intelligence T.J.M. Bench-Capon, Paul E. Dunne Department of Computer Science, University of Liverpool, Liverpool, United Kingdom Received 27 April 2007; received in revised form 27 April 2007; accepted 1 May 2007 Available online 10 May 2007 Abstract Over the last ten years, argumentation has come to be increasingly central as a core study within Artificial Intelligence (AI). The articles forming this volume reflect a variety of important trends, developments, and applications covering a range of current topics relating to the theory and applications of argumentation. Our aims in this introduction are, firstly, to place these contributions in the context of the historical foundations of argumentation in AI and, subsequently, to discuss a number of themes that have emerged in recent years resulting in a significant broadening of the areas in which argumentation based methods are used. We begin by presenting a brief overview of the issues of interest within the classical study of argumentation: in particular, its relationship in terms of both similarities and important differences to traditional concepts of logical reasoning and mathematical proof. We continue by outlining how a number of foundational contributions provided the basis for the formulation of argumentation models and their promotion in AI related settings and then consider a number of new themes that have emerged in recent years, many of which provide the principal topics of the research presented in this volume. 2007 Elsevier B.V. All rights reserved. Keywords: Argumentation models; Dialogue processes; Argument diagrams and schemes; Agent-based negotiation; Practical reasoning 1. Introduction In its classical treatment within philosophy, the study of argumentation may, informally, be considered as concerned with how assertions are proposed, discussed, and resolved in the context of issues upon which several diverging opinions may be held. Thus philosophical investigations of argumentation, from Aristotle to the present day, have addressed such themes as: the mechanisms by which legitimate argumentation in support of a claim may be distinguished from flawed argumentation; analyses of the typical structures that constitute argument components and argumentation development; the processes by which participants engaging in debate may advance their respective positions and undermine contrary stances and arguments, etc; and the contexts in which these questions are decided. The importance of such philosophical theories to so-called everyday reasoning has a long and distinguished history in AI, and contributions from contemporary philosophical analyses continue to play a major role in the evolution of effective computational exploitation of argumentation technology. Within the simplified overview of argumentation outlined in the preceding paragraph, one can, already, identify a number of themes whose elements embody issues of a computational nature in the following: * Corresponding author. E-mail address: ped@csc.liv.ac.uk (P.E. Dunne). 0004-3702/$ see front matter 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.artint.2007.05.001
620 T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 Defining the component parts of an argument and their interaction. Identifying rules and protocols describing argumentation processes. Distinguishing legitimate from invalid arguments. Determining conditions under which further discussion is redundant. It is, of course, the case that similar issues underpin one well-established and highly-developed theory: that of formal logic and mathematical proof. It is no coincidence that much of the formal computational treatment of argumentation has its roots in ideas developed from AI inspired contributions to logic and deductive reasoning. So one finds in mathematical proof theory core concepts such as: precisely defined means for expressing assertions (e.g. formulae in a given logical language); accepted bases on which to build theorems (e.g. collections of axioms); procedures prescribing the means by which further theorems may be derived from existing theorems and axioms (e.g. templates for inference rules); and precise concepts of termination (e.g. a sentential form is derivable as a theorem, true ; or is logically invalid, false ). While the structural elements presented in this view of mathematical reasoning have proven to be a useful basis in the development of argumentation-based models in AI, the formal apparatus and methods of mathematical reasoning are, ultimately, radically different in nature to those of importance when considering the concept of argumentation as it is familiar from everyday contexts, e.g. as it might occur in political debate, the discussion of ethical principles, deliberation in judicial settings, etc. While there are, of course, parallels that can be made, e.g. that those engaged in debate have some collection of accepted premises on which there is agreement, possibly, even, some recognition of when contributions to a discussion are unreasonable or flawed, etc. there are, however, a number of fundamental distinctions between the concepts P is a formal proof that T holds and P is a persuasive argument for accepting T. Thus, in mathematical reasoning, (a) The premises can, ultimately, be explicitly defined in terms of closed concepts, e.g. the axioms of Euclidean geometry, the Zermelo Frankel basis for set theory (ZF). Furthermore classical mathematical reasoning is based on an assumption that such premises are, collectively, consistent. 1 (b) Reasoning and analysis takes place within a closed, tightly defined context, i.e. there is no notion of incomplete or uncertain information. (c) Conclusions are final and definite: if P is a correct proof that T, then T is, ipso facto valid and this status does not admit subsequent qualification or amendment, let alone retraction. (d) Reasoning and conclusions are entirely objective, not susceptible to rational dispute on the basis of subjective views and prejudices. 2 Proof is demonstration whereas argument is persuasion. In argument and discussion as encountered in everyday contexts, it is rare that any, let alone all, of these apply: the premises upon which debates may build are often presupposed as forming part of the background assumptions common to all parties involved; the information and knowledge brought to bear in the course of discussion will often be incomplete, vague, or uncertain. The remaining two aspects, in many ways, highlight the most significant differences between logical proof and persuasive argument. Arguments are defeasible: the reasoning that formed a persuasive case for T, in the light of changes in viewpoint or awareness of information not previously available, may subsequently fail to convince. This defeasibility is never removed: an argument may cease to be challenged and so accepted, but the possibility of challenge remains. Finally, the extent to which an argued case is accepted is subjective, dependent on the views, attitudes, and prejudices of the audiences to which it is directed. The same case may convince some people but, equally, fail to convince others. 1 We note that in a number of systems, consistency cannot be formally proven, cf. [95] and so, in such cases, consistency is, indeed, an assumption. 2 Some clarification of this claim may be in order. Suppose is a derivation of ϕ within a theory A,R (with axioms A and inference rules R). Within the same theory, the proof admits no rational, objective basis for dispute: criticisms that ϕ is inconvenient or counter-intuitive are subjective, and entirely irrelevant to the status of ϕ within the theory A,R. In order to give rational grounds for not accepting ϕ it is necessary to endorse an alternative theory within which ϕ cannot be derived. As a concrete example, consider the axiomatic basis ZF extended by the so-called Axiom of Choice (ZF+ AC): although widely adopted in modern theory this conflicts with Intuitionist principles which disqualify AC asanaxiom so that theorems dependent on AC are (rationally) not accepted by Intuitionists.
T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 621 One can summarise the distinction between argumentation and proof by the observation that the object of argumentation is to persuade (to acceptance of a given claim; to performance of a desired action, and so on). Unlike the concept of proof at the level of deriving a sentential representation of an assertion whether an argument is correct is not a factor, and, indeed, correctness may not even be sensibly defined. In contrast, mathematical reasoning, in order to have any value, must be correct where correctness has a strict, formal definition: beyond this requirement, however, notions of persuasiveness are unimportant. In summary, the importation of elements from logic and formal deductive reasoning has provided a powerful basis for modelling and analysing argumentation in computational settings of AI. As we shall discuss later, these continue to form an important strand of contemporary work. It is also the case, however, that a number of significant directions pursued in recent years, have broadened the scope and concerns of argumentation in AI beyond this earlier logic driven motivation. As a consequence, one has a shift of emphasis within the developed treatment of argumentation in AI progressing from formalisms rooted in classical deductive reasoning through models handling concepts of incomplete information and uncertainty, to precise semantics for capturing defeasibility, and, within recent work, propounding computational bases to account for subjective aspects of argumentation, often using the notion of audience introduced by Perelman [145]. One consequence of such analyses has been the growth of work dealing with computational procedures and issues of resource-boundedness in implementing these, e.g. as discussed in Loui [118]. All of these features make argumentation particularly attractive to applications requiring distributed intelligence, autonomous components and synchronous interaction. To conclude this overview, it is worth noting one further historical and continuing tradition. In common with many established areas of AI, the computational theory of argumentation has benefited from contributions and ideas originating in many diverse disciplines, so that a number of fundamental themes draw on the earlier work of, for example, philosophers, logicians, and legal theorists. In presenting concrete realisations of these theories, work on argumentation in AI has in its turn informed further research in these fields. Argumentation is thus an excellent example of interdisciplinary interchange and the mutual benefits that can stem from this. 2. Foundations of argumentation in AI A discussion of early influences on the development of argumentation models in AI may be found in the comprehensive survey of Chesnevar, Maguitman, and Loui [57], so we will be content merely to outline a few significant aspects, referring the reader to [57] for a more detailed exposition. We concentrate on three important influential themes, 2.1 Origins in non-classical logic. 2.2 Models of argumentation as dialogue process. 2.3 Diagrammatic views of argument structure. 2.1. Influence of non-classical logics on argumentation in AI Early studies using argumentation inspired methods in AI contexts can be found in the work of Birnbaum, Flowers, and McGuire, [42] in which a structural model of argument embracing notions of support and attack within a graphtheoretic base comprising propositional forms, is applied to textual reasoning; and Alvarado and Dyer s approaches, [4,5], to the analysis of editorial presentation. Undoubtedly, the important early motivations that brought argumentation theory into use in AI arose from the issues of reasoning and explanation in the presence of incomplete and uncertain information. The failings of classical propositional logic as a means to address these had been delineated in the influential work of Reiter [165], and, a pressing concern of work throughout most of the 1980s and early 1990s was to build on the proliferation of treatments of non-monotonic logics within AI. This state is succinctly summarised by [57, pp. 337 338]. Within AI, several non-monotonic reasoning formalisms emerged...in these formalisms, conclusions drawn may be later withdrawn when additional information is obtained. Formal logics of argument emerged as one style of formalising non-monotonic reasoning. The literature on non-monotonic reasoning dominated AI s journals in the mid 1980s.
622 T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 Thus argumentation was initially adopted as a possible supporting approach with which to effect a formal treatment of non-monotonic reasoning, rather than as a paradigm whose study might be of independent interest in itself. The engagement of philosophers and legal theorists with reasoning and argumentation in AI marked a key stage in the move towards computationally grounded models of argument. Particularly notable is the impact of Pollock s work on defeasible reasoning and justification: originally promoted in specialist philosophical literature, e.g. [146 148] its relevance and significance to AI was recognised following Pollock s dissemination of these ideas in [149 151]. In parallel with this development of the formal logical theory in which context the significance of argumentation techniques with respect to non-classical logic was further emphasised in the contributions of Simari and Loui [173] and Brewka [44] (ideas in the latter being subsequently developed in [45]) the early 1990s saw important uses of argumentation techniques in the computational treatment of legal reasoning: notably in Rissland and Ashley s treatment of legal argumentation from Case Law, [14,167,174] and its later extension by Aleven [3]; Prakken s analyses in [152]; Sartor s models of legal reasoning as described in [169,170]; the use of argumentation techniques in explaining complex legislation from Bench-Capon, Coenen, and Orton [29], etc. The technical treatment evident in AI contributions to non-monotonic logics and the argumentation-based methodologies offered in the field of legal reasoning found some degree of common ground in the exploitation of logic programming paradigms and knowledge-based systems. It was in this context, building on argument-based treatments of negation-as-failure of Kakas, Kowalski, and Toni [108], together with Eshghi and Kowalski s work on abductive interpretation [89], that the watershed contribution of Dung [72,73] appeared: the model of argumentation described in [73] is now recognised as providing an important bridge between argumentation theory as a supporting analytic tool for non-monotonic reasoning and the independent exploitation of argumentation models in wider AI contexts. Two important ideas are put forward and expanded in [73]: (A) The reduction of argumentation about a given issue to a completely abstract setting consisting of a set of atomic arguments, X, and a binary relation over these, A X X, with x,y A interpreted as the argument x attacks the argument y. (B) The proposal that intuitive notions of collection of justified arguments can be formally described by that of extension-based semantics: that is, through various properties of subsets, S of X within an argumentation framework (AF), X, A. 3 The effect of (A) is that neither the structure of an argument nor the nature and semantics underpinning x attacks y need explicit consideration within the abstract framework. Thus an argument, x, may be a simple atomic proposition, p; or a (defeasible) rule, e.g. p q r; or an instantiation of a richer, more particular, perhaps even domain specific, argument scheme. That x attacks y may be on account of reasons varying in form from x promotes a claim logically equivalent to the negation of that promoted by y, e.g. x : p and y : p; or x promotes a claim incompatible with the premises supporting the claim in y, e.g. x : p and y : q p r, and so on to the extent that attacks disputing the applicability of a given inference scheme and more complex structures are represented entirely abstractly in a single binary relation. Dung s introduction of various extension-based semantics has, as we shall discuss in Section 3.1, had a profound influence on subsequent analyses of the concept of collection of justified arguments. In extremely informal terms, an extension semantics, E, can be thought of as describing properties that a subset of arguments within a given framework must satisfy in order to be deemed collectively justified, i.e. E : X, A 2 X {, }. Dung demonstrates how different choices of E may be used to colour varying degrees of an argument s acceptability ranging from very liberal (so-called credulous) conditions through to extremely restrictive (so-called sceptical) requirements. The elements of Dung s original set-theoretic semantics are reviewed in a number of articles in this issue and for further technical exposition we direct the reader to the article in this volume by Baroni and Giacomin [22]. The past 5 7 years have witnessed an intensive study of mechanisms with the common aim of developing Dung s ideas in various directions. For a detailed comparative critique of abstract argumentation techniques we refer the reader to the valuable perspective provided by Vreeswijk [182]. 3 The significance of this work is enhanced since approaches which include additional information, such as preferences, may do so in such a way that the evaluation of argument status remains in terms of an underlying abstract framework.
T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 623 Subsequent work [43] of Dung in conjunction with Bondarenko, Kowalski and Toni, makes explicit the link between abstract argumentation and uniform treatment of non-classical logics. The Assumption-based frameworks (ABF) of [43] consider deductive theories L, R (with L a formal language such as the language of well-formed propositional sentences, and R a (countable) set of inference rules) augmented by a triple T,Ab, in which T L is a set of beliefs, Ab L a (non-empty) set of assumptions and : Ab L maps assumptions to their contrary in L. 4 Such frameworks are shown to be applicable as a generic approach to describing a wide range of non-classical logics 5 including: Reiter s Default Logic [165], Moore s Autoepistemic Logic [131], logic programming, and divers other non-monotonic reasoning formalisms. While ABF structures may on first inspection seem unrelated to the abstract argumentation frameworks of [73], these can be presented as AFs by building the attack relation from the contrary mapping. A fuller overview of this approach may be found in, e.g. the paper of Dung et al. in this volume [76, Section 2.2]. One feature of importance in this abstraction of ABFs is that the resulting structure will typically describe an infinite graph. Informally, extensionbased semantics for ABFs are introduced as subsets of Ab whose union with the belief set T constitute a consistent theory. Fuller technical descriptions may be found in [43, pp. 70 71]. Such links between abstract argumentation frameworks and the deductive bases underpinning assumption-based schemes bring two powerful analytic approaches to bear in algorithmic studies of extension-based semantics for argumentation: combinatorial and algorithmic graph theory have been usefully applied in the former case; whereas technology developed for deductive reasoning and formal logic has provided insight into the latter. We expand on such computational and algorithmic issues in Section 3.1. 2.2. Argumentation and dialogue processes The perspective of argumentation presented in Section 2.1 is strongly biased to a view wherein the overall aim of argumentation is in deciding the status of some claim and in presenting a justification for it: thus, an assertion, p, is established in the light of available information, but recognised as potentially defeated should new data emerge; the nature of justification often being through some logical reasoning process. In total such a view treats argumentation in support of a claim as a somewhat one-sided process in which a single party merely presents a reasoned justification. In many applications such an abstraction is, of course, a natural analogue to use, e.g. in explanation-driven systems such as [29], or in the context of decision-support processes. An objection to such treatments, however, is that they fail to embrace the dialectical nature of argument, discourse, and debate as encountered in everyday contexts: here argumentation is rarely a matter of a single party presenting a case but is more commonly an informed exchange of ideas and positions involving several contributors: in other words, argumentation concerning an issue, typically, arises as a dialogical process. Given this it is, perhaps, surprising that significant computational exploitation of the established treatments of dialogue within philosophical, rhetorical, and linguistic analyses, has been a comparatively recent phenomenon. Although originally explored to a limited extent as a means of interacting with expert systems, the significant factor motivating contemporary computational use of dialogue methods can be found in supporting multi-agent system applications, a topic that we review in Section 3.2. In this section we review a number of foundational contributions and the preliminary AI motivated developments of these. As with the developments discussed in Section 2.1 many of the ideas within computational treatments of dialogue build on contributions originating from philosophical analysis. One established concern of such study is the notion of fallacy : a key aspect of which is the view that so-called fallacious argument encompasses a much wider collection of issues than simply what may be (more accurately) termed erroneous (mathematical or logical) reasoning. Thus, argument employing fallacious reasoning (in this wider philosophical and rhetorical sense) is not ipso facto wrong nor easily dismissed merely by the action of highlighting occurrences of fallacy. 6 It is the case, however -and here one finds a basis for the interaction between argumentation and dialogue processes that particular fallacies occasion 4 Some treatments of ABFs omit explicit specification of a belief base T, e.g. [75,76]. 5 Choosing L, R to be the language and standard inference structures of classical propositional logic, together with ϕ = ϕ that is the contrary is simply logical negation the resulting ABF structures recover standard propositional reasoning. 6 Although it is, of course, true that in regarding errors of logic as a form of fallacy, indicating such incidences could suffice as an attack on the argument in whose support they are used.
624 T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 potential attacks on a specific line of argument. For example, argumentum ad verecundiam whereby a claim is unreasonably justified through appeal to the opinion of an authority, is regarded in rhetorical theories of debate as fallacious. The use of such support cannot be properly deemed invalid merely by signalling its occurrence. Such fallacies are, however, open to attack by processes that invite further discussion. Thus the argument, ϕ is the case since X has stated as much can be attacked in several ways, e.g. by disputing the authority of X in matters relating to the domain of ϕ, by challenging the assertion that X has made any statement regarding ϕ, etc. We, thus, find two important themes emerging: the classification of distinct types of fallacy; and the nature of possible attacking arguments. Categories of fallacious reasoning have long been a topic of interest in rhetorical analysis and the treatment of Hamblin [100] has had some influence on computational ideas. An important contribution to the second issue the nature of attacks on fallacious argument is found in Walton s formulation of Critical Question [186] which has been adopted in several computational treatments of persuasive argument, e.g. [17,18,155,188]. An influential contribution in which a number of key ideas that have played a significant role in computational realisations of dialogue machinery, also arose in modelling and detecting one specific type of fallacious reasoning: MacKenzie s dialogue game, DC, aimed at exposing uses of petitio principii in argumentation, [121]. A striking feature of MacKenzie s analysis, 7 evident from [121, Appendix, pp. 129 132], is the wealth of computational ideas that are introduced and their expression in an operational form. Thus, the concepts of commitment store, dialogue rules, locutions, etc. have all been adapted and extended in current dialogue-based applications. Among the earliest argumentation-based presentations of MacKenzie s ideas beyond their use in recognising a specific class of fallacious reasoning, are those of Moore [130] and Bench-Capon, Dunne, and Leng [31]. Analyses of fallacy in debate contexts had, typically, not differentiated the range of styles and aims to which dialogue processes might be directed. This concept that dialogues could be distinguished by their intentions is expanded in the seminal work of Walton and Krabbe [189]. The taxonomy of dialogue types in [189] is neither intended to be nor presented as a definitive, complete catalogue of dialogue forms, 8 but it has had considerable influence on the treatment of dialogue in multi-agent systems. The central significance of [189] to later work on argumentation in AI is in promoting an awareness that the purposes of dialogue encompass a number of different aims and, therefore, the appropriate procedural mechanisms (for example, as might be defined from MacKenzie s model) employed in computational use have distinctive requirements, e.g. the operational specification of dialogue processes geared to negotiating agreements are unlikely to be best-suited to use in dialogues whose purpose is to elicit information. Dialogue game approaches figure in several influential contributions dating from the mid-late 1990s: Gordon s Pleadings Game, [96]; Lodder s study of legal justification (Dialaw) in [113]; and Loui s use of dialectic approaches to non-monotonic reasoning in which one of the first considerations of computational limits is presented [118]. A currently active area in which these ideas have proven to be highly relevant is the exploitation of argumentation in multi-agent systems applications. We conclude this discussion of dialogue processes by reviewing one further aspect that has been fruitfully adapted to argumentation in AI: analysing argument justifiability via dialogue games. Interpretations of mathematical reasoning as a dialogue process have been advocated, from the early 1960s, in work of Lorenz and Lorenzen [115 117]. An approach that has provided a useful abstraction within the more general context of argumentation considers discussion over a disputed argument, p, as involving two participants conventionally denoted PRO (who argues in favour of p) and OPP (who objects to p). A generic dialogue game building on Dung s argumentation model is presented in Jakobovits and Vermeir [105] (see also [104, Section IV]) and this approach has been used as the basis of a number of later studies, e.g. [30,51,52,70,71]. This view of argument justification as resolved by a dialogue implicitly underpins a number of formal ideas that have been adopted in algorithmic methods, e.g. the concepts of argumentation lines and the generalisation of such as argumentation proof-trees. Analysis of the properties of such structures has proven useful in examining a number of issues in argumentation and we discuss such approaches further in Section 3.1. 7 It should be noted that [121] appeared almost 30 years ago in 1978. 8 Indeed, independently Dunne, Doutre, and Bench-Capon [85] and Walton, himself, in [187] have analysed one dialogue form not presented in [189].
T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 625 2.3. Diagrammatic treatments of argument structure In the presentation above we have made a distinction between the concepts of argument and argumentation: the latter being understood as the processes by which given arguments are analysed and evaluated. An argument may, informally, be considered as the basic supporting case behind a given assertion. Thus an argument, for p, in this sense, may itself give rise to a variety of distinct structures ranging in complexity from simple statements of fact ( p is an accepted fact ), through to deductive templates (e.g. p follows from q and q is the case ) to even more intricate structures (or argument schemes) that themselves may rely on further (sub)-arguments. A variety of argument schemes have been proposed and studied, in work by legal theorists and philosophers, e.g. [145,186], however, the bulk of our preceding review has largely addressed the issue of argumentation and its evolving application within AI with the notion of argument itself only briefly discussed. It has been seen that Dung s fundamental model, as described in [73], abstracts away such internal structure from individual arguments in order to focus on the manner in which arguments interact via the defined attack relationship. In unfolding the exact nature of the argument x attacks the argument y, however, the reason why such an attack is present needs to be considered in terms of those structural schema underlying the arguments x and y from which the attack arises. Such an interpretation, therefore, raises issues that concern the form an argument might take, i.e. issues regarding the components and representation of arguments rather than the process and outcome of the argumentation involved. Questions regarding argument form and uniform treatments of these have assumed increasing importance in recent years, especially with respect to multi-agent exploitation of argumentation methods. A key aspect of this work has been the extent to which diagrammatic models of argument structure have been adopted. Early diagram-based models of argument were intended to aid in illustrative hand construction and analysis of argument, with the resulting schemes being static depictions. An important example of such an approach is provided by Wigmore diagrams [190], which although used as a method of describing legal arguments have only recently been rediscovered and promoted in AI contexts, e.g. [41,164]. More widely known is the highly influential model of argument promoted by Toulmin in the 1950s [178]. Toulmin s structural interpretation treats an argument as consisting of five sub-components: the Claim advanced (which could be qualified by a modal operator to describe concepts such as normally ); the (factual, evidential etc.) Data supporting this Claim; a Warrant providing a licence to infer the Claim from the Data, together with Backing for this Warrant; and, to encapsulate exceptional cases, Rebuttal conditions. Although in common with Wigmore s scheme Toulmin diagrams were originally presented as a static representation of the totality of an argument, they have proven a flexible approach in AI treatments of argumentation. Thus, Bench-Capon et al. [32] describe a dynamically evolving extension of Toulmin s schema and its use in a dialogue game. Later work of Bench-Capon [26] develops the dialogue game of [32] providing a complete move repertoire and operational semantics for it. The exploitation of such argument diagram techniques offers an important basis for a number of contemporary ideas among which are: argument visualisation methods, e.g. as might be used in decision support and explanation; argument construction from source material; the specification of methods for interchanging arguments between distinct parties; and in providing a unifying link between informal argument descriptions and formal abstract approaches such as [73]. Current work in these areas will be discussed in Section 3.4. 3. Recent trends and concerns Section 2 has offered a, necessarily condensed, summary of influences on argumentation in AI covering up to the start of the present century. As we turn now to more recent developments, that is subsequent to those contributions discussed in [57,182] a number of trends becomes apparent: the continuing enrichment of the formal theory of argumentation building on [43,73,105]; the growth of argumentation-based methodologies in multi-agent systems applications; new computational treatments of argument diagramming and visualisation; the exploitation of argumentation in novel specialist domains; and the development of theoretical bases embracing subjectivity in argumentation and concepts of practical reasoning. Overall one finds in such themes a broadening of the scope of argumentation in AI beyond its earlier traditional uses in realisations of non-classical logic scenarios. In this section, we discuss some of these themes in greater depth with particular reference to the articles contributing to this volume.
626 T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 3.1. Development of the Dung-style model of argumentation The graph-theoretic model of argumentation framework in [73] and the deductive schema supporting the assumption-based frameworks of [43] have given rise to an extensive body of research with particular concentration on the following, (a) Extension based semantics of argumentation. (b) Algorithmic and complexity issues in argumentation. (c) Dialogue processes for deciding acceptability. Extension-based semantics of argumentation Each of the extension-based semantics presented in [73] builds on sets of arguments, S, that are conflict-free: that is, no argument within S attacks another in S. Conflict-freeness, as observed by Baroni and Giacomin [22] in their study of evaluative criteria for extension semantics, is viewed as a minimal requirement to be satisfied within any computationally sensible notion of collection of justified arguments. Conflict-freeness, however, is too weak a condition, in itself, to be applied as a reasonable guarantor that a set of arguments S is collectively acceptable : for example, such a set could be attacked by arguments not among its members. It is in the approaches posited to form additional conditions on (conflict-free) subsets of arguments that complications become apparent, and out of the divers methods proposed to resolve such complications that the current, to use Guillermo Simari s phrase, plethora of argumentation semantics 9 has emerged. The three principal extension-based semantics introduced in [73] the so-called Grounded, Preferred and Stable semantics can exhibit a variety of problematic aspects. 10 (P1) Emptiness: although an extension satisfying the prescribed conditions always exists, there are AFs for which the only such extension is the empty set. This can arise with both the grounded and preferred semantics of [73]. (P2) Non-existence: an extension, when it exists is never empty, but there are frameworks for which no extension meeting the required criteria exists. This can occur, for example, with Dung s stable semantics. (P3) Multiplicity:inanAF there may be several incompatible extensions, i.e. sets S 1 and S 2 which are well-defined extensions of X, A but with S 1 S 2 failing to be so. While Dung s grounded semantics does not suffer from this problem, frameworks are easily constructed in which both the preferred and stable semantics exhibit this phenomenon. A number of approaches have been proposed in order to address these and other perceived drawbacks. Thus, Cayrol and Lagasquie-Schiex [55] define concepts of graduality in order to evaluate classes of acceptable arguments, Caminada [46] introduces semi-stable semantics; Dung, Mancarella and Toni [75, p. 151] develop ideal semantics and their paper in this volume [76] presents further analyses concerning the computation of ideal extensions in ABFs. Baroni et al. [21,23] define various extension-based semantics for an AF building from the strongly-connected component (SCC) decomposition of its directed graph: of the resulting SCC-recursive semantics, CF2-semantics have been examined in depth in [23]. In [62], Coste-Marquis, Devred and Marquis consider a refinement of the concept of conflict-free set in order to exclude controversial arguments [73, p. 332], i.e. arguments {x,y} such that, although x,y / A there is an indirect attack by x on y: the resulting approach gives rise to the prudent semantics of [62]. A number of extension-based semantics have been proposed motivated by new interpretations of the interactions between arguments that should be considered: thus the basic binary attack relation of [73] is developed. Important contributions of this type include the articles by Cayrol et al. [50,53,54] wherein the relation the argument x supports the argument y is introduced leading to the formulation of bipolar argumentation frameworks. In such frameworks each of the existing extension semantics can be qualified through bipolarity, e.g. [50] considers bipolar prudent semantics. Other developments of Dung s attack structure are offered in work of Nielsen and Parsons [134] 9 During the presentation of [129] at COMMA 2006, 12th September, 2006. 10 It should be noted that although our description is given in terms of AFs exactly the same issues arise in the analogous semantics within ABFs.
T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 627 using an approach predicated on the idea that a binary attack relation is not always appropriate and thus this should be defined in terms of A 2 X X, i.e. each x X has an associated set of subsets of X that attack it. 11 A different treatment of A forms the basis of Amgoud and Cayrol s approach in [7] to the interpretative issues created by the presence of multiple preferred extensions. In this any AF is augmented by a preference relation over arguments defining an attack in A. In the resulting preference-based AF (PAF) 12 an attack x,y is relevant only if the argument y is not preferred to the argument x. By imposing suitable restrictions the effect of specifying preferences within a given AF is to reduce it to an acyclic graph: for such frameworks, [73] has shown that the grounded, preferred, and stable semantics coincide in a unique extension. The value-based argumentation frameworks (VAF) of Bench-Capon [27,28] also stem from attempts to provide a formal basis on which to rationalise choices between several preferred extensions. The basic elements of VAFs are described in the paper by Dunne [80, Section 8]. In common with the preference-based approach, resolving choices in VAFs can be interpreted in terms of consistently removing attacks (using value orderings) so that the resulting framework is, again, one in which all three basic extension semantics coincide. The philosophical rationale underpinning VAFs ultimately derives from Perelman [145] and is treated in more depth in Section 3.3. A detailed comparative discussion of preference and value-based methods may be found in [30, Section 7.1]. The article by Baroni and Giacomin [22], makes a powerful case for re-examining the proliferation of new semantics: In fact, various kinds of motivations have been used to support the introduction of new semantics with respect to classical proposals... These motivations range from the desire to formalise some high-level intuition, not captured by other proposals, to the need to achieve the correct treatment of a particular example (or family of examples) regarded as particularly significant.... Clearly, these kinds of heterogeneous intuitions hardly lend themselves to systematic comparisons. Given this situation it is not surprising that comparisons are quite often carried out using specific problematic examples, often ingeniously devised so as to bring to light patently different behaviours exhibited by the semantics under discussion. ([22, Intro.]) Extension-based semantics in AFs continues to be an extremely active topic for argumentation models in AI and a number of specialised technical questions remain unresolved. 13 Important as such questions are, it may well be the case, however, that treatments of extension-based semantics will come to focus less on the construction of novel specialised forms and more on consolidation theories such as the evaluative principles of [22] or the complementary approach applied to arguments with a particular structure of [47,48]: just as the attempts to construct a notional definitive non-monotonic logic from the disparate alternatives proposed in the 1980s are now recognised as illfated, such is likely to be the outcome of efforts to build an ultimate extension-based semantics. Algorithmic and complexity issues While the preponderance of formal theoretical study into computational issues arising from [43,73] has addressed semantic concerns within these abstract frameworks, there is a significant core of results relating to algorithmics and computational complexity. Early work of Dimopoulos and Torres [69] derived exact complexity classifications for a number of decision problems involving extension-based semantics in AFs 14 and a summary of these results may be found in [80, Table 1(a d)]. In [81], Dunne and Bench-Capon further develop complexity-theoretic analysis of Dung s model in deriving exact bounds on the computational complexity of two questions (neither of which is considered in [69]): that of deciding if a given argument is justified under the most restrictive semantics defined in [73] (so-called sceptical acceptance); 11 For details see the paper by Nielsen and Parsons [135] in this volume. 12 These should not be confused with the partial argumentation frameworks (also denoted PAF) described in the article by Coste-Marquis et al. [64] in this volume. 13 Among which are issues such as conditions under which particular extension-based semantics coincide, existence properties etc. 14 The analyses of [69] are not presented in the context of Dung s frameworks from [73] but may be readily translated into this. A discussion of the links between [69] and [73] may be found in Dunne and Bench-Capon [81, pp. 188 189].
628 T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 and the problem of determining whether an AF satisfies Dung s concept of coherence [73, Definition 31(1), p. 332] whereby every set of arguments defining a preferred extension also defines a stable extension. Both [69] and [81] relate to decision problems in AFs: the basis for hardness results is via suitable directed graph mechanisms, cf. the two core constructions described in [80, Secn. 3]. In an important series of articles [66 68] Dimopoulos, Nebel and Toni consider analogous questions within various instantiations of assumption-based frameworks. A significant achievement of this work is in characterising the computational complexity of decision questions in ABFs with respect to that of testing derivability (i.e. of ϕ from a given base ) within the associated logic modelled by the ABF: derivability being central in determining the existence of attacks between arguments. In consequence a number of instantiations of ABFs describing particular non-classical logics exhibit a significant increase in complexity compared with the NP or p 2 -completeness of related problems in AFs. The contributions of [66 69,81] primarily focus on purely complexity-theoretic analyses. Efficient algorithmic methods are introduced in [73] for special classes of AF (directed acyclic graphs DAGs) and, more recently, in work of Coste-Marquis et al. [63] (for symmetric frameworks). The extent to which such graph-theoretic conditions can ameliorate complexity issues forms the central topic of Dunne s paper [80] in this volume. The treatment of graduality in [55] includes a number of algorithmic elements and other useful work has emerged from modelling of argument justification via dialogue games, e.g. [51,52,184]. One collection of methods which have received increasing attention over the past five years concern enumerative techniques for constructing all extensions (of a particular form) within a given AF, key contributions being the work of Doutre and Mengin [70] and Verheij s labelling approach to generating all stable extensions of an AF described in [181]. More recent work includes algorithms of Nielsen and Parsons [134] and Vreeswijk [183], the first of these relative to the authors set-theoretic notion of attack mentioned earlier. An alternative slant on the question of enumerating preferred sets is offered in [77] which considers the following issue: under the assumption that an enumeration, S, has already been produced, to what extent can it be represented compactly with the representation allowing S?S to be decided efficiently for any subset of arguments of S? 15 While [77] presents indications that a number of computational question remain difficult even with significant additional information provided, more positive results indicate that concise encodings of preferred sets which can be efficiently queried may be possible. 16 The algorithmic analysis of ABFs has been rather less advanced than that within AFs: one major factor accounting for this is, of course, the formidable complexity-theoretic issues raised in [68]. Nevertheless, promising dialogue based techniques are presented in work of Dung, Kowalski and Toni [74]. The paper by Dung et al. in this volume [76] offers a further example of dialogue approaches by adapting these to the computation of ideal extensions in ABFs. In addition, recent work of Egly and Woltran [88] is of some interest: this proposes approaches building on translations to quantified Boolean formulae and the subsequent exploitation of highly-tuned QBF solvers to resolve decision questions. 17 The article by Coste-Marquis et al. in this volume [64] introduces an important topic that appears to have been largely neglected in previous studies: given a number of distinct AFs, describing, say the views of a number of observers regarding a specific issue, how should these be merged into a sensible unified framework that fairly reflects individual viewpoints? The techniques in [64] contribute both to semantic and algorithmic aspects of this question. For the developments of [73] represented by PAFs and VAFs, only the latter gives rise to non-trivial algorithmic and complexity issues. Treatments of these, including exact complexity classifications of the principal decision questions together with algorithmic approaches may be found in the series of papers by Dunne et al. [30,71,83,84]; Dunne s paper in this volume indicates that a number of non-trivial issues remain to be resolved in the algorithmic treatment of VAFs. Dialogue-based approaches to deciding argument acceptability The view of reasoning as a dialogue mechanism has been widely adopted in formal algorithmic approaches to determining the acceptance status of arguments within both AF and ABF models. Such a view has also featured 15 While the AF itself provides a compact encoding of its preferred sets, in view of [69] it is unlikely that this would satisfy the efficient querying criterion. 16 Of course, finding such encodings given an AF is another matter. 17 Treatments of non-classical logic via propositional encoding had also been proposed in earlier work of Ben-Eliyahu and Dechter [25]; similar techniques, for AFs, are discussed in [81, p. 202].
T.J.M. Bench-Capon, P.E. Dunne / Artificial Intelligence 171 (2007) 619 641 629 significantly in models of argument methods building on deductive reasoning templates. Key ideas underpinning these techniques include argument line (a chain x 0,x 1,...,x k in which the argument x i attacks the argument x i 1 for i>0) and the concept of (partial) proof tree, which, in informal terms, can be interpreted as combining a number of distinct argument lines concerning a common initial argument. In pursuing such approaches a number of basic questions arise: rules and strategies affecting the selection of arguments with which to continue a dialogue; termination properties; demonstrating soundness and completeness of procedures intended to establish acceptability of arguments in particular semantics; approaches to assessing the efficiency of dialogue methods, and so on. The generic formalism for describing dialogue games within Dung s model of argument introduced in [105] was discussed, briefly, earlier. A significant subsequent development is found in the methods presented by Vreeswijk and Prakken in [184]. This describes the structure of Two-part Immediate Response (TPI) disputes. Adopting the two player PRO and OPP convention for debate over an argument x, among the features of TPI-disputes is the requirement for each player to attack the most recently played argument of their opponent whenever it is possible (within the game s rules) to do so. Several examples in [184] establish that both players require moves allowing back-tracking to a defined earlier point in a dialogue. The resulting game is shown to be sound and complete for so-called credulous reasoning, i.e. where the aim is to decide if x is a member of at least one preferred extension. Thus for any AF and argument x within it, TPI-disputes are guaranteed to terminate and correctly to determine whether x is justified under Dung s credulous preferred semantics (PRO wins) or x cannot be so justified (OPP wins). A variant of this game provides sound and complete methods for sceptical reasoning in coherent AFs i.e. where PRO wins if and only if x belongs to every preferred extension. Vreeswijk and Prakken s results in [184] were instrumental in motivating one of the first systematic studies concerning formal concepts of efficiency of dialogue games in argumentation: the definition and analysis of dispute complexity presented by Dunne and Bench-Capon in [82]. Informally, the dispute complexity of a dialogue game is measured in terms of the (worst-case) number of moves that might be required in order to resolve the status of a given argument in an AF. One significant contribution of [82] is its positioning of such dialogue games within an established body of work regarding the relative efficiency of propositional proof methods via the concept of dispute complexity, i.e. the basis provided in Cook and Reckhow [61]. Thus, [82] not only demonstrates that TPI-disputes occasion a propositional proof method but also, adopting the comparative criteria for such systems presented in [61], further show that the resulting system is equivalent to the CUT-free sequent calculus of Gentzen [92]. In consequence, via results of Urquhart [180], one may construct (a family of) AFs and arguments within these AFn,ϕ such that resolving the status of ϕ requires exponentially long TPI-disputes. Important treatments combining elements of MacKenzie s dialogue model [121] with the formal approach of [105] for example, locutions, utterances, rules for dialogue continuation, termination are found in McBurney, Parsons, and others, e.g. [106,122 124,141,142]. Much of the emphasis of this work is directed towards providing a basis for dialogue exploitation in multi-agent system contexts for example Torroni s analysis of termination properties in negotiation dialogues [177] as discussed in Section 3.2. Treatments of proof-theoretic techniques via dialogue methods using the concepts of argumentation line and partial proof-tree have been considered in a number of recent papers. An important issue in this context concerns the design of heuristics that reduce the search space thereby obviating the requirement to consider all expansions of each argument line. Dung, Kowalski and Toni [74] propose a novel backward reasoning approach to the construction of prooftrees in ABFs. Recent work of Chesnevar and Simari [58] deals with sceptical argumentation via a lattice-theoretic encoding of the relevant search space. A related question in implementing dialogue mechanisms is that of deciding which (from a range of available options) is the best continuation for a participant to contribute. There are, of course, many interpretations of best that may be applicable from loosely defined intuitive qualitative notions (e.g. most persuasive or convincing ) to quantitative ideas, e.g. guaranteed to terminate debate in the fewest possible moves. In [86,87] Dunne and McBurney consider one formalisation of this problem that allows it to be related to the literal selection problem examined by Liberatore [112]. A final collection of issues, concerning which computational treatment has only recently been initiated, addresses questions arising from wider considerations of the motives of participants. Thus, recognising that contributors to a discussion may have rational bases to obstruct its development or be anxious to avoid revealing information regarding their pursuit of an issue. In [91], Gabbay and Woods examine the use of so-called stone-walling tactics as one means of impeding the progress of information-seeking dialogues, while Dunne [78] considers settings in which one participant seeks to prolong discussion and reviews such approaches against a variety of legal applications. Informa-